Gaussian Blur implementation not exactly working

1
votes

I am trying to implement Gaussian Blur from scratch in c. I have a program that creates a struct and loads a .bmp file, creates a Gaussian blur filter (the first two for-loops below), then sends the struct with height, width, and actual pixels info to another function that applies the filter to the image (the 4-nested for-loops below). After it is applied, it outputs the modified .bmp file. Currently it slightly works, but it's not working all the way. I can't adjust the intensity of the blur (all the levels of blur look the same). It also appears to miss the very outer edge of pixels all around the image.

When I'm creating the filter, I use:

double sigma = 10.0;
double r, s = 2.0 * sigma * sigma;
double sum = 0.0;

But it looks the same whether sigma is 1, 10, 100 - my understanding was that it should get blurrier with bigger sigma.

creating the filter:

// generate 5x5 kernel
for (int x = -2; x <= 2; x++)
{
    for(int y = -2; y <= 2; y++)
    {
        r = sqrt(x*x + y*y);
        gKernel[x+2][y+2] = (exp(-(r*r)/s))/(M_PI * s);
        sum += gKernel[x + 2][y + 2];
    }
}

// normalize the Kernel
for(int i = 0; i < 5; ++i)
    for(int j = 0; j < 5; ++j)
        gKernel[i][j] /= sum;

and then applying the filter:

for (i = 1; i < bmp->height - 1; i++)
{
    for (j = 1; j < bmp->width - 1; j++)
    {
        sum = 0.0;
        for (p = 0; p < 5; p++)
        {
            for (q = 0; q < 5; q++)
            {
                sum += bmp->pixels[(i + p) * bmp->width + j + q] * gKernel[p][q];
            }
        }
        bmp->pixels[(i - 1) * (bmp->width) + j] = (unsigned char) sum ;
    }
}

Any help much appreciated! Thank you for your time.

2
cgaussianbmp
Welcome to Stack Overflow. Following the SO conventions makes it more likely you'll be helped. Please read how to provide a MCVE. stackoverflow.com/help/mcve - Gene
Are you increasing your filter kernel size when you increase sigma? - Gavin Portwood
No - you mean generating a x by x kernel for a sigma of y, and then a x+5 by x+5 kernel for a sigma of y+5? I have only tested for a kernel size of 5 by 5. Should I increase that value when I increase sigma? - kab403

2 Answers

2
votes

It misses the points round the edge because you have told it to do so. It is also hard-coded to a kernel of width 5 pixels.

Edges are always a problem with filtering. Generally it is best to expand the image, padding with mirroring. Then apply the filter without special code for the edge condition. The width * height * N * N method is very slow. For a big filter, you needs to use an FFT for filtering. That's a much more advanced technique.

0
votes

I figured this out earlier today - thanks to those who responded, y'all were a lot of help! Basically once I added in code to increase the size of my kernel as I increased the sigma, I was able to adjust how blurry the image appeared. I also added code to handle the edge cases - on that note, I am currently using a if-else if-else statement, but am trying to figure out padding instead.

Now I have:

// generate nxn kernel
for (int i = 0; i < kSize; i++)
{
    for(int j = 0; j < kSize; j++)
    {
        gKernel[i][j] = exp(-(i*i+j*j)/(2*sigma*sigma))/(2*M_PI*sigma*sigma);
        sum += gKernel[i][j]; 
    }
}

And you can imagine all '5's above are replaced now with 'kSize'. When I apply the filter, I adjust the for loops to go from 0 to bmp->width/height and then adjust where I place the pixel inside the body of the loop (this is where the if-else if-else statement comes into play, though I'll change that soon). Thanks again!